SCEC2025 Plenary Talk, Seismology

Enhanced earthquake detection with graph neural networks: Applications to northern California seismicity

Ian W. McBrearty

Oral Presentation

2025 SCEC Annual Meeting, SCEC Contribution #14382
Developing improved earthquake catalogs is an essential goal in seismology, where enhanced catalogs can substantially improve our view of active seismogenic processes, fault network distribution, and velocity models of the Earth. However developing accurate low-magnitude-of-completeness catalogs is challenging due to the large abundance of seismic picks and available stations. These issues are compounded by high event rates, high false pick rates, time-varying station networks, and other forms of noise and artifacts corrupting the data. Here I review recent graph neural network-based approaches for earthquake phase association (GENIE) and double-difference relocation (GraphDD), where GNNs are used to determine the seismic source history from continuous streams of picks, and to relocate large (> 100,000’s) earthquake catalogs that span diverse seismogenic regions. I review the basic strengths of GNNs for robustly handling high volumes of noisy data and heterogenous, time-varying station and source distributions, as well as the particular design choice of introducing both a station and a source graph into our GNN architecture. Through the use of graph message passing, which shares information between neighboring stations/sources, while simultaneously accounting for the relative distances and positions between objects, the model is more stable and adaptable compared to non-graph (or Transformer) models. The utility of both techniques is demonstrated in the construction of a long-term, deep learning enhanced catalog of northern California (> 2 million events) that shows high spatial resolution and consistency with the existing catalog, while increasing the detection rate by ~4x. Quality control and analysis of the catalog indicates promising initial results, and highlights areas for future improvement, such as identifying systematic errors in the initial picks, and improving robustness of the models by training on more realistic synthetic data.